Open Access

ARTICLE

A Method for Detecting Spatio-Temporal Correlation Anomalies of WSN Nodes Based on Topological Information Enhancement and Time-Frequency Feature Extraction

Miao Ye¹, Ziheng Wang¹, Qiuxiang Jiang¹, Xingsi Xue², Wenxi Liu³, Yu Ning¹, Cheng Zhu^1,4,*

1 School of Information and Communication, Guilin University of Electronic Technology, Guilin, China
2 Fujian Provincial Key Laboratory of Big Data Mining and Applications, Fujian University of Technology, Fuzhou, China
3 College of Computer and Data Science, Fuzhou University, Fuzhou, China
4 Information Center, Guilin Medical University, Guilin, China

* Corresponding Author: Cheng Zhu. Email:

Computers, Materials & Continua 2026, 88(2), 77 https://doi.org/10.32604/cmc.2026.078282

Received 28 December 2025; Accepted 28 April 2026; Issue published 15 June 2026

Abstract

In recent years, anomaly detection in Wireless Sensor Networks (WSNs) has been widely studied using Graph Neural Networks and Transformer-based methods. However, in multi-node and multi-modal data scenarios, these approaches still face challenges such as insufficient extraction of spatiotemporal correlation features, limited modeling capabilities when relying solely on either time-domain or frequency-domain information, and high computational overhead. To address these issues, this work aims to develop an anomaly detection model that balances detection performance with computational efficiency, enabling effective identification of complex anomaly patterns. Specifically, we propose a time–frequency feature extraction method with topological information enhancement, topology-enhanced multi-modal spatio-temporal anomaly detection (TE-MSTAD). Building upon the Receptance Weighted Key Value (RWKV) model with linear complexity, a cross-modal feature extraction module is introduced to strengthen the modeling of multi-modal correlations. Meanwhile, adaptive adjacency matrices are constructed by integrating time–frequency features and combining outputs from different Graph Neural Networks, thereby enhancing topological information. Furthermore, a dual-branch structure is designed to jointly model time-domain and frequency-domain features, improving the extraction of complex anomaly characteristics. Experiments on both publicly available datasets and real-world collected data demonstrate that the proposed method achieves F1-scores of 92.52% and 93.28%, respectively, outperforming existing methods in detection performance and generalization capability.

---

Keywords

---